Ghost in the Droid

Control real Android and iOS devices, run apps, take screenshots, and execute on-device inference via Ghost in the Droid MCP server (62 tools).

Ghost in the Droid is an MCP server that gives any LLM agent a real Android or iOS device as its body. 62 tools covering tap, swipe, type, screenshot, screen-tree reading, app launch, camera, TTS, crash reports, batched execution, and on-device inference.

Install: pip install ghost-in-the-droid Repo: https://github.com/ghost-in-the-droid/android-agent

When to use this skill

Use Ghost in the Droid when the user wants to:

• Control an Android phone or tablet via ADB
• Control an iPhone or iPad via WebDriverAgent
• Take screenshots of a real device screen
• Tap, swipe, type, or scroll on a device
• Launch or interact with mobile apps
• Run on-device inference (llama.cpp, MediaPipe, MLX)
• Automate mobile workflows end-to-end
• Spin up Android emulators via Docker+KVM pools

Setup

1. Install the MCP server

bash

pip install ghost-in-the-droid

2. Add to your MCP config

For Claude Code (~/.claude/claude_code_config.json):

json

{
  "mcpServers": {
    "ghost-in-the-droid": {
      "command": "python",
      "args": ["-m", "ghost_in_the_droid"]
    }
  }
}

For OpenClaw:

bash

openclaw mcp add ghost-in-the-droid --command "python -m ghost_in_the_droid"

3. Android prerequisites

bash

# Install ADB
sudo apt install android-tools-adb   # Linux
brew install android-platform-tools  # macOS

# Connect device (USB or TCP)
adb devices

4. iOS prerequisites

WebDriverAgent must be installed and running on the target device. See the iOS setup guide.

Core tool reference

Screen interaction

Tool	What it does
tap	Tap at coordinates or element
swipe	Swipe between two points
type_text	Type text into focused field
press_key	Press a key (HOME, BACK, ENTER, etc.)
scroll	Scroll up/down/left/right
long_press	Long press at coordinates

Screen reading

Tool	What it does
screenshot	Capture device screen as image
screen_tree	Get UI accessibility tree (for clicking elements by label)
find_element	Find element by text, id, or class
get_screen_text	OCR-extract all visible text

App management

Tool	What it does
launch_app	Launch app by package name
close_app	Close app
install_apk	Install APK from path or URL
list_apps	List installed apps
clear_app_data	Clear app storage

Device info

Tool	What it does
list_devices	List connected ADB devices
device_info	Get device model, OS, battery, etc.
get_clipboard	Read device clipboard
set_clipboard	Write to device clipboard

On-device inference

Tool	What it does
run_llm	Run inference via llama.cpp on-device
run_mediapipe	Run MediaPipe model on-device
run_mlx	Run MLX model (iOS/macOS)

Emulators (Docker+KVM)

Tool	What it does
create_emulator	Spin up new Android emulator
list_emulators	List running emulator pool
destroy_emulator	Tear down emulator instance

Batching

Tool	What it does
run_flow	Execute a sequence of actions as one atomic batch
run_skill	Replay a saved skill (recorded interaction)

Usage patterns

Take a screenshot and describe what you see

text

1. Call screenshot → get base64 PNG
2. Pass to vision model for description

Find and tap a button by label

text

1. Call screen_tree → get accessibility tree
2. Find element with target label
3. Call tap with element bounds

Automate a login flow

text

1. launch_app("com.example.app")
2. screenshot → verify login screen
3. tap(username_field)
4. type_text("[e-mail]")
5. tap(password_field)
6. type_text("password")
7. tap(login_button)
8. screenshot → verify logged in

Multi-device targeting

Always use device_id parameter to target a specific device when multiple are connected:

text

tap(x=100, y=200, device_id="emulator-5554")
screenshot(device_id="R3CN20ABCDE")

Tips

• Always call list_devices first to confirm devices are connected
• Use screen_tree over coordinate-based tapping when possible — it survives resolution changes
• run_flow is much faster than individual tool calls for scripted sequences
• For iOS, WebDriverAgent must be already running on the target device before using any iOS tools
• On-device inference tools require the model to be pre-loaded on device; call run_llm with a load action first